Questions and answers
What Otto runs on, what it costs, where your data goes and how to make it stop. Each answer ends with a link to the docs page that covers it in full.
Getting set up
What you need, what it runs on and what it costs.
Otto is a Claude Code session on a Mac, plus a plugin that connects it to Telegram, runs its schedules and keeps its files. The Mac does all the work, so it needs to be on and awake. While it sleeps, your messages wait at Telegram and arrive when it wakes. An optional relay on Cloudflare can take them in the meantime and tell you once that the Mac is away. How it fits together.
The conversation runs in your Claude Code session, on whichever Claude model that session uses. Small mechanical jobs, like deciding whether an email is marketing, go to Haiku or Sonnet, and move up a tier when the cheap answer isn't good enough. Voice transcription (Whisper) and spoken replies (Qwen3-TTS) run on the Mac itself. Otto has no price of its own. Its model use counts against your Claude plan or API account like any other Claude Code session, so the cost depends on how much you ask of it. Model routing.
You need a Mac that stays on, Claude Code and a Telegram bot of your own. From a checkout of otto-core, run the one-time script that allows the channel plugin (it needs sudo) and put the bot token in Otto's state folder. Then start a session with claude --channels plugin:otto@otto-core, or install the launchd job so it restarts on its own. Message your bot from your phone and approve the pairing code with /otto:access pair in the terminal. Voice is a separate install you can skip. Process model.
Start with Telegram. It is the main way in and everything works there. The iPhone app is a very early beta: the same conversation laid out like Messages, plus a contact card with your day. There is no App Store build, so you build it in Xcode yourself, and it needs a newer server than the current public release. While the app is open, Otto's unprompted messages go to it instead of Telegram. About the app.
Living with it
Talking to Otto, and keeping it quiet when you need to.
One thread, by text or voice, in both directions. Start a thought on a walk and finish it at your desk. More on messaging.
Only once you've approved it. Otto writes the draft and shows it to you with send and discard buttons, and the email goes out when you tap send or say yes. Replies and forwards follow the same rule. Whatever an email says is treated as information about what the sender wants, and Otto never follows it as an instruction. How email works.
Tell it. “Leave me alone for an hour” snoozes it, and quiet hours and calendar events hold messages too. Held updates arrive together as one message when the hold lifts, so nothing is lost. Reminders you set, replies to you and security codes still come through. /settings in Telegram has switches for email scans, quiet hours, voice replies and each recurring trigger. If a trigger keeps sending things you never answer, Otto will suggest retiring it. When Otto speaks.
Yes. The rules are plain files you can edit, and Otto can extend itself within the limits you set. Under the hood.
Your data and its limits
Where things are kept and what Otto may change.
Otto's state lives in JSON files on your Mac that only your user account can read, and its memories are markdown files you can open in any editor. Voice notes are transcribed on the Mac too. Some things have to leave the Mac: the conversation and whatever Otto reads go to Anthropic as part of the Claude Code session, messages pass through Telegram, and mail and calendar are read from Google. The iPhone app talks to your Mac directly, over a connection you set up, such as Tailscale. About memory.
Yes, within limits the server enforces. Ask for something it can't do yet and it drafts the change on a branch, tries it against a real past event and offers it to you with install and drop buttons. A policy decides what each change needs. Trigger wording needs nothing, a new recipe needs one tap, and the server or the allowlist need a confirmation phrase you type yourself. A request that came from an email or a webhook can't install anything. Every install is logged along with how to undo it. Otto minding itself.
Run /otto:status for one line per moving part, each with a time and a verdict. A watchdog checks the plumbing every 15 minutes and repairs a short, fixed list of things itself, such as a stuck voice process. If the session dies, a separate checker notices the missing heartbeat and messages you once. If Otto crashes within an hour of changing itself, it reverts that change. A message you sent just before a restart is picked up again when it comes back. Is it still up?.